Complex engineered systems arise throughout computing, communications, and networking. Many factors, each having a finite number of levels, impact the behaviour of the system either singly or in interaction with one another. Testing or evaluating such a system involves formulating a set of tests, when executed, responses or outcomes from the tests are analyzed. A single round of testing is conducted. To witness the effect of an interaction, some test must cover it, this does not suffice in general to locate the interaction or to measure its effect. When there are few factors or many tests, experimental designs can measure (and hence locate) the interactions. When there are many factors and few tests, can we locate the interaction(s)? Can we efficiently detect them?Combinatorial arrays, locating and detecting arrays, are introduced to address such location and detection in the context of combinatorial testing. Locating and detecting arrays are contrasted with covering arrays and with experimental designs. An application to a 75 factor protocol stack for file transfer is given to demonstrate their practical use. Finally, their place in the literature of combinatorial testing is discussed and some directions are outlined.
Today, commercial load balancers are often in use, including in the production network at Arizona State University (ASU). One of the main issues such load balancers face is that they use a static scheme for load distribution. However, at particular times of the academic year, such as during course registration, the network exhibits significant variations in both temporal and spatial traffic characteristics. At these times, students experience much greater latency and become frustrated with the network service. To address this problem, our aim is to develop an SDN-based approach to load balancing to better cope with the traffic variation.
Wireless systems exhibit a wide range of configurable parameters (factors), each with a number of values (levels), that may influence performance. Exhaustively analyzing all factor interactions is typically not feasible in experimental systems due to the large design space. We propose a method for determining which factors play a significant role in wireless network performance with multiple performance metrics (response variables). Such screening can be used to reduce the set of factors in subsequent experimental testing, whether for modelling or optimization. Our method accounts for pairwise interactions between the factors when deciding significance, because interactions play a significant role in real-world systems. We utilize locating arrays to design the experiment because they guarantee that each pairwise interaction impacts a distinct set of tests. We formulate the analysis as a problem in compressive sensing that we solve using a variation of orthogonal matching pursuit, together with statistical methods to determine which factors are significant. We evaluate the method using data collected from the w-iLab.t Zwijnaarde wireless network testbed and construct a new experiment based on the first analysis to validate the results. We find that the analysis exhibits robustness to noise and to missing data.